The Week Your Agent Bill Could Double

One week. Two pricing shocks. And the most important open-weights release of the year.

If you run AI agents for a business, this was the week the meter changed. Here's what happened.

Featured: DeepSeek Just Rewrote the Price Sheet

DeepSeek launched V4-Pro on Aug 14 and raised API prices by up to 1,100%. The old flat rates are gone, replaced by peak and off-peak pricing:

• Output tokens: 6 yuan → 27 yuan per million during peak hours (9am-noon, 2pm-6pm Beijing time). That's the 1,100% jump.
• Uncached input: 3 yuan → 9 yuan per million. Cached input: 0.025 → 0.3 yuan.
• Off-peak is half the peak rate — so your bill now depends on what time of day your agents run.

DeepSeek claims V4-Pro beats xAI's GLM-5.2 on several agent benchmarks and beats Claude Opus 4.8 in selected tests, though it trails Kimi K3 and Claude Fable 5 in most evaluations. On Artificial Analysis it scores 53 — tied with GLM-5.2.

The business lesson isn't about DeepSeek specifically. It's that no API price is permanent. A model you've built your pipeline around can change its economics overnight — and your automation either absorbs it or breaks.

"Every agent stack that depends on one provider's price sheet is a liability. This week proved it twice."

Google's Counterpunch: Gemini 3.7 Flash at Half Price

Google shipped Gemini 3.7 Flash three weeks after 3.6 Flash — and the numbers jumped hard:

• FrontierCode 1.1: 34.4% → 43.6%
• DeepSWE v1.1: 49% → 65.3%
• AutomationBench: 17% → 30.4% — nearly doubled for agentic workflows
• 1M-token context, Terminal-bench 2.1 at 85.8%

The catch: intro pricing of $0.75/M input and $3.75/M output runs through Dec 31, then doubles to $1.50/$7.50 on Jan 1. Same intro-pricing cliff, different vendor. If you're evaluating Flash-class models for production, the eval window is now — while the cheaper rate is still live.

The Open-Weight Answer: Qwen 3.8 27B

Alibaba's Qwen team dropped Qwen 3.8 27B under Apache 2.0 — a 27B model with integrated vision, 262K native context (extensible to 1M), and a Gated DeltaNet architecture. The FP8 variant scores:

• 73.0 on Terminal-Bench 2.1 — agent benchmark
• 84.3 on OSWorld-Verified — computer use
• 61.7 on SWE-Bench Pro, 90.3 on LiveCodeBench v6

The HN thread hit 1,090 points in a day. A 27B model you can self-host with vision and 262K context, under a permissive license, is exactly the escape hatch API price hikes need. Your agents don't have to live on someone else's price sheet.

Quick Hits

• Hermes Agent v0.20.1 (Aug 13) — Nous Research's patch release rolls up ~656 merged PRs since v0.20.0. The open-source agent keeps shipping weekly.
• Anthropic: $11.5B Q2 revenue, first profitable quarter — 14x jump over Q2 2025's $787M, more than double Q1. The agents business became a real business.
• OpenAI enterprise revenue overtakes ChatGPT consumer — CFO Sarah Friar says the 60-40 consumer-enterprise split has crossed. Agents are now the growth engine.
• Skan AI raised $63M — process intelligence company up 300% YoY, grounding agents in how work actually gets done. The "agents in production" wave keeps funding.

Tool of the Week: Qwen 3.8 27B

Worth a weekend experiment. Grab the FP8 weights, run it locally (it's 27B — a serious GPU or a quantized build), and point one low-stakes workflow at it. Even if you stay on API models for production, you've now got a fallback that no vendor can reprice. The prices will keep moving; your infrastructure shouldn't have to.

The Takeaway

DeepSeek raised prices 1,100%. Gemini's intro rate doubles in January. Anthropic and OpenAI posted numbers that make agents look like the core business. The model market is moving fast in both directions — and the only stable strategy is to not be married to any single price sheet.

Wednesday's deep dive: the model-agnostic playbook — five rules for building agents that survive price hikes, plus how to run your own evals while intro pricing lasts.

Automate & Grow with A.I.
Written by Michael Devellano — building AI automation systems for founders and agencies.
Subscribe for weekly AI automation insights